Measuring River Wetted Width From Remotely Sensed Imagery at the Subpixel Scale With a Deep Convolutional Neural Network

Abstract River wetted width (RWW) is an important variable in the study of river hydrological and biogeochemical processes. Presently, RWW is often measured from remotely sensed imagery, and the accuracy of RWW estimation is typically low when coarse spatial resolution imagery is used because river boundaries often run through pixels that represent a region that is a mixture of water and land. Thus, when conventional hard classification methods are used in the estimation of RWW, the mixed pixel problem can become a large source of error. To address this problem, this paper proposes a novel approach to measure RWW at the subpixel scale. Spectral unmixing is first applied to the imagery to obtain a water fraction image that indicates the proportional coverage of water in image pixels. A fine spatial resolution river map from which RWW may be estimated is then produced from the water fraction image by superresolution mapping (SRM). In the SRM analysis, a deep convolutional neural network is used to eliminate the negative effects of water fraction errors and reconstruct the geographical distribution of water. The proposed approach is assessed in two experiments, with the results demonstrating that the convolutional neural network‐based SRM model can effectively estimate subpixel scale details of rivers and that the accuracy of RWW estimation is substantially higher than that obtained from the use of a conventional hard image classification. The improvement shows that the proposed method has great potential to derive more accurate RWW values from remotely sensed imagery.